The free layer of your stack turned out to be the expensive one

⚡ 60-Second Catch-Up

Three of the things you build on got bought this week. Nvidia is reported to be taking Hugging Face — The Information put it at $12.9B agreed, Business Insider reported talks above $13B, and Reuters said flatly that it could not verify either. AWS is taking DuckLabs, the team behind DuckDB. Stripe is taking Clerky, which handles the paperwork behind 23% of all Silicon Valley seed and pre-seed financings. None of these are products you pay much for. All of them are load-bearing.

The pattern is not "big company buys small company." It is that the free, neutral layer of the stack turned out to be the most valuable real estate on it, and neutrality is the thing being purchased. Hugging Face reportedly turned down a $500M Nvidia investment at roughly a $7B valuation specifically to stay neutral; the reported number now is nearly double that. DuckLabs says DuckDB, DuckLake and Quack stay open source and the team stays in Amsterdam — which is the right promise to make and the one every acquired open-source project makes.

→ do this: for each of these, find out today whether you depend on the licence or on the hosted service. An Apache-2.0 file on disk is yours forever. A free API key, a free tier, and a default registry endpoint are not. The difference takes ten minutes to check and is the whole of your exposure.

Meanwhile, everyone is trying to stop paying Nvidia — including the company buying Hugging Face. OpenAI published its first real benchmarks for Jalapeño, its custom inference chip: roughly 3.4× lower end-to-end latency and 1.5× more peak throughput per kilowatt than a GB300 on Kimi K2.5, at half the package power. Apple's M6 and M5 Ultra, out earlier this week, push the same work onto the desk. Perplexity and Nvidia launched a fully local agent with no per-token cost. Applied Compute launched a cloud built specifically for open weights.

The take: inference is decoupling from training, and it is decoupling fast. Jalapeño cannot train a model — OpenAI still buys Nvidia for that. But the expensive, repetitive, forever part of running an AI product is inference, and that part is about to get cheap in ways that are not evenly distributed between whoever has silicon and whoever has an API bill.

→ do this: stop hard-coding one provider. If your product calls a model, put a router in front of it now, while switching is a config change rather than a rewrite. The teams who can move workloads are the ones who will capture this.

And the productivity numbers still refuse to show up. McKinsey's State of AI 2026 has company AI adoption at 89%, up from 21% in 2017, with 80% of workers saying AI made them more productive — and company-wide profit impact stuck at 37%, unchanged year over year. That gap is the actual story of 2026.

The one counter-example worth reading closely is Dee Kapila's account of how Fin's customer-success org went AI-first: the data team shipped a packaged, pre-installed Claude setup, gave the GTM team unmetered access with no token rations, ran a two-hour hands-on class, and measured adoption by depth of use rather than logins. Someone on the team published 100 builds. She automated a two-hour Sunday reporting ritual down to a click.

→ do this: note what is missing from that story — no procurement cycle, no pilot committee, no per-seat rationing. The enablement was two hours and the constraint removed was permission. If your AI rollout has a budget but no unmetered access, you have bought the 37% number.

Tools

Build & ship

• ⭐ BrowserOS neo — An open-source browser your agents drive using your own logged-in sessions — one-click import from Chrome, then hand Claude Code, Cowork, Codex or Cursor the tab. Replay any session to see what it actually did. · Free, open source
• MulmoTerminal — Runs many Claude Code and Codex sessions in one browser terminal grid, and shows you which agent is waiting on you. Local, tmux-backed. · Open source (MIT)
• whip — A coding-agent harness in Go: tool-use loop, TUI, provider-routable models with live catalog discovery, MCP support, background subagents. · Open source (Apache-2.0)
• Coldtea — Agents that test every pull request on a real device, watch production, and file the bugs they find before your users do. macOS. · Free to try
• Tailcat — Like netcat, but over Tailscale's data plane and without its control plane. · Open source (BSD-3)
• OpenTag — A self-hosted AI on-call triage bot for Slack and Teams, shipped as a fork-and-go starter app. · Open source (MIT)

AI & agents

• Applied Compute AC2 — A training and inference cloud built specifically for open weights — the internal platform Applied Compute uses for Microsoft, Nvidia, Cognition, Mercor, DoorDash and Harvey. · Book a demo
• JoyAI-Echo — JD's open release for long audio-visual generation. · Free (GitHub)
• Nitro — Professional human translation behind an API your agent can call on its own — no account, no API key, 80+ languages, payment negotiated over a 402 challenge. · Pay per request

Design & create

• ⭐ Nebula Sans — A humanist sans in two styles and six weights, built as a drop-in alternative to Whitney SSm. Based on Adobe's Source Sans, released under the SIL Open Font License. · Free (SIL OFL)
• Grainient — 375 grainy gradients, 406 smooth ones, 90 animated, 538 AI backgrounds, plus a shader tool. · Paid (Pro)
• Limora — Upload your logo, colours and fonts once; generate on-brand images, backgrounds, OG cards and product shots across 14 asset types. · Paid (credits)
• blokdots — Hardware prototyping for designers — wire up physical components with if-this-then-that rules, no code. Pro tiers add standalone deployment and JavaScript/Arduino export. macOS. · Free, paid Pro

Growth & ops

• ⭐ Accept: text/markdown — A convention plus a checker for serving a Markdown variant of your pages to AI agents via content negotiation, so crawlers spend context on your prose instead of your DOM. Run any URL through the scorecard. · Free
• Is GitHub Cooked? — Filters GitHub's whole incident history down to the services and severities you actually depend on, so "GitHub is unreliable" becomes a number you can argue with. · Free
• Risklytics — Insurance built for AI-agent and robotics companies — tech E&O, cyber, D&O — plus plain-language guides to the insurance clause in your customer contracts. · Quote

Resources

Steal the template

• Discovery Meetings Playbook — a full founder-sales discovery script as a working doc you can copy and edit. 28 minutes to read, considerably less to steal.

Learn the craft

• How to evaluate LLMs before production — GitHub's own method for deciding whether a model is good enough to ship behind.
• Agentic engineering for reliable coding agents — practical guardrails for getting repeatable output from coding agents.
• Card Sorting and Tree Testing — how to get information architecture right before you build the nav.
• Where Should AI Go in Your UI? — a placement guide for AI features, which is now a design problem rather than a model problem.
• Granite 4.2 LLMs: How They're Built — IBM walks through the training and architecture decisions in the open.
• Balancing flex items with flex-wrap — Ahmad Shadeed on the layout case CSS still makes awkward.

Benchmarks you can re-run

• Jalapeño's first results — OpenAI publishes the method alongside the numbers: workload, model, sequence lengths, package TDP. Worth reading as a template for how to report a benchmark you want believed.

Reads

☕ Under 5 minutes

• Vocab Break — Ian Barber (4 min)
• Python's pre-declared constants are kinda weird (4 min)
• When str.lower() is a security vulnerability in Python — Seth Larson (3 min)
• New domain for Sign in with Apple — if you ship Sign in with Apple, this one is a to-do, not a read (1 min)
• One Go binary, one YAML file, one SQLite database: why I wrote my own monitoring tool (5 min)

🍵 5–10 minutes

• How Universities Should Prepare Founders — Paul Graham (9 min)
• You need to find product-market fit again (sorry) — PostHog (7 min)
• Why Bootstrapped Businesses Are More Relevant Than Ever — Crunchbase News (6 min)
• Building May Be Cheap Now. Being Wrong Still Isn't — Viget (6 min)
• How AI is actually remaking security (9 min)
• Compounded Mastery (9 min)
• Robotics companies should go full-stack (8 min)
• No shortcuts: why AI threatens creative instinct — It's Nice That (6 min)
• The Hidden Debt of Convenience — Sahil Bloom, on inconvenience as a tax on exposure to reality, and what optimising it to zero costs you (9 min)

📚 Longer, worth it

• The 3 New Pricing Models in B2B — SaaStr, on why per-seat is dying and what replaces it (14 min)
• OpenAI and Anthropic Could Dominate Global AI Compute — Dwarkesh Patel with Dylan Patel (48 min)
• Compression is prediction — ngrok (24 min)
• Dark Mode Toggles: Two States Are Enough — Lea Verou (10 min)
• Max Mullen on when to take advice from your investors (12 min)
• Short-Lived Credentials for AI Agents — Vercel, on what replaces the long-lived API key (12 min)
• Let's build your great helper — Jonathan Ward on why an assistant loyal to you has to be one you own (17 min)
• Why performant code matters (but gets widely ignored) — Gergely Orosz interviews Casey Muratori on profiling as a local minimum, reading assembly, and his case against test-driven design (1h 53m)

Numbers

• $12.9B — what The Information reported Nvidia has agreed to pay for Hugging Face. Business Insider separately reported talks above $13B; Reuters said it could not verify either report.
• $4.5B — Hugging Face's valuation at its 2023 Series D, which raised $235M in a round led by Salesforce Ventures.
• ~$7B — the valuation implied by a roughly $500M Nvidia investment offer Hugging Face turned down earlier, to protect its neutrality (Business Insider).
• 23% — share of all Silicon Valley seed and pre-seed financings accounted for by startups using Clerky, which Stripe has agreed to acquire (Clerky).
• $140B+ — aggregate venture capital raised by companies formed on Clerky (Clerky).
• 6.5× — how much faster startup formation on Clerky grew in the past year than its historical average (Clerky).
• 89% — companies using AI in 2026, up from 21% in 2017 (McKinsey, State of AI 2026).
• 37% — share of companies reporting a company-wide profit impact from AI, unchanged from last year, against 80% of workers who say it made them more productive (McKinsey).
• $30T — the total addressable market Anthropic is preparing to put in front of IPO investors, ahead of SpaceX's $28.5T claim from its May filing (WSJ).
• $2.4T — combined earnings last year of the 1,500 largest US public companies, for scale against that TAM slide (FactSet).

What happened

Models & math

• OpenAI published the first benchmark results for Jalapeño, its custom inference chip: against a GB300 on Kimi K2.5, roughly 3.4× lower end-to-end latency, 3.8× lower minimum time-between-tokens, and 1.5× higher peak throughput per kilowatt — at 700W package power versus 1,400W. It runs models; it cannot train them, so OpenAI still buys Nvidia for training (OpenAI).
• Perplexity and Nvidia launched Portable Computer, a fully local AI agent with no per-token cost, positioned as the application that finally makes Nvidia's DGX Spark worth owning (VentureBeat).
• GLM-5.3-Flash shipped from Z.ai and Qwen3.8-Flash-Next from Qwen, both aimed at the cheap-and-fast end of the market.
• Applied Compute launched AC2, a training and inference cloud for open-weight models, naming Microsoft, Nvidia, Cognition, Mercor, DoorDash and Harvey as customers (Applied Compute).

Safety & policy

• OpenAI published a postmortem on the Hugging Face incident, along with a technical report (OpenAI).
• Sign in with Apple is moving to a new domain. If you have it in production, this is a change you have to make rather than one you can read about (Apple Developer).

Business moved

• The Hugging Face sale process appears to have found its buyer. Last week the company was working with a bank to gauge interest at $13B or more, with nothing agreed. The Information now reports an agreed deal with Nvidia at $12.9B; Business Insider reported talks above $13B; Reuters said it could not immediately verify. Nothing has been confirmed by either company (The Information, Business Insider, Reuters).
• AWS is acquiring DuckLabs, expected to be effective in early September. Mark Raasveldt and Hannes Mühleisen say the team stays together in Amsterdam and that DuckDB, DuckLake and Quack remain open source (DuckLabs).
• Stripe has agreed to acquire Clerky, the startup-formation and financing-paperwork service used by a large share of Silicon Valley seed rounds (Clerky).
• Amazon Mechanical Turk closes permanently on September 30, 2026 (Amazon Mechanical Turk).
• OpenAI's head of data centers, Chris Malone, left last week — another senior departure as the company heads toward an IPO and ramps data-centre spending. He joined in March 2025, just after Stargate was announced (WSJ).
• Anthropic is preparing to tell IPO investors its addressable market tops $30 trillion, which would beat SpaceX's own record claim (WSJ).

Platforms shifted

• Webflow is now available inside Codex and ChatGPT (Webflow).
• Anthropic merged memory across Claude chat and Claude Cowork, on by default, so context carries between them (The Next Web).
• ChatGPT can now generate custom iMessage and WhatsApp stickers (9to5Mac).
• EVE Online has begun its migration to Python 3 — two decades of production Python, moving (CCP Games).

One note on what we left out: we skipped a handful of front-page stories with nothing in them for founders, and cut one "AI founding team" product whose landing page carried testimonials describing a different product entirely.